基于无人机交通视频多尺度特征融合的对象级车辆行为分析方法研究

Dynamic monitoring of transportation operation and the security analysis of driving behavior serve as vital foundations for improving transportation efficiency and reduce traffic accidents. Traditional traffic monitoring methods suffer from their fixed location, small coverage range, vulnerability to occlusion and low reliability. Hence, they cannot fully meet the needs of modern smart transportation systems. Compared with traditional data acquiring methods, the low-altitude remote sensing videos captured by unmanned aerial vehicle (UAV) can span over a vast area of the transportation scene with various angles. With an ultra-high resolution, UAVs can record much more detailed information of various traffic objects. However, the UAV videos also pose significant challenges to existing traffic video processing methods. In this study, we focus on busy road sections of megacities and capture high resolution traffic videos using UAVs. We then develop new vehicle detecting and type identification methods based on the deep convolutional neural network, and combine it with the particle filtering based vehicle tracking and the double spectral clustering based trajectory modeling algorithms. We will use these techniques to build spatial-temporal vehicle driving behavior analysis methods, with an aim to effectively eliminate various noises contained in the vehicle shadows of the remote sensing videos while improving the integration and efficiency of relevant algorithms. Our work helps to develop the low-altitude remote sensing ultra-high resolution video processing theory, and inspire new ideas and strategies for designing transferrable and efficient vehicle behavior analysis methods. Besides, our work exhibits high-level theoretical value and promising potentials for industrial applications.

交通运行的动态监测及驾驶行为安全分析是提升交通效率、减少交通事故的重要基础。传统交通监控手段因安设位置、覆盖范围、遮挡干扰、和可靠性等局限，难以满足现代智慧交通系统的建设需求。新兴基于无人机平台的低空遥感交通视频监控技术，较传统手段数据获取视角更优、覆盖范围更广、分辨率更高，包含的交通对象也更多，但也对传统交通视频分析算法提出了新的挑战。本项目利用无人机采集大城市重点监控交通路段的交通视频，设计新的基于多尺度深度卷积神经网络的车辆检测、车型识别模型并将之与经典的基于粒子滤波的多目标跟踪算法和基于二次谱聚类的轨迹建模算法有机结合，建立时空一体化的车辆行为分析方法，以期解决低空遥感视频中噪声繁杂、车辆阴影干扰严重、传统算法集成度和效率低下等关键科学问题。该研究有助于推动超高清低空遥感视频处理理论，为实现泛化能力强、计算效率高的车辆行为分析提供新思路和新方法，具有积极的理论研究和工程应用价值。

交通运行的动态监测及驾驶行为安全分析是提升交通效率和减少交通事故的重要基础。传统交通监控手段因安设位置、覆盖范围、遮挡干扰、和可靠性等局限，难以满足现代智慧交通系统的优化管理需求。新兴基于无人机平台的低空遥感交通视频监控技术，较传统手段数据获取视角更优、覆盖范围更广、分辨率更高，包含的交通对象也更多，但也因其观测点多变、尺度不一、干扰噪声多等因素对传统交通视频分析算法提出了新的挑战。本项目提出了基于sparse-RCNN的车辆精准检测与准确车型识别算法，该方法抛弃传统密集型网络的锚点框等内容，直接学习“稀疏”候选范围，实现车辆检测与车型识别。本项目还提出了一种基于交并比的车辆轨迹提取方法，该方法利用相邻两帧之间车辆的重叠区域提取车辆轨迹，可快速准确地实现大数据量的车辆跟踪和轨迹提取。提出了基于改进碰撞时间的车辆冲突预测模型，在实时提取的车辆轨迹的基础上，引入车辆尺寸信息，对传统车辆碰撞模型进行改进，有效地提升了车辆冲突的检测的准确性。提出改进社会驱动力模型的行人运动轨迹模型算法，通过在考虑自驱动力、边界驱动力和周围行人驱动力的基础上，引入自行车和汽车对人运动的影响，并根据最大似然法计算各个驱动力对行人运动的影响。经与行人真实轨迹分析对比，证明所提出方法的精度较高。通过该项目系列研究的开展，有效地推动了基于无人机视频的对象级车辆行为分析方法研究进展。